Slow feature extraction and Wasserstein distance adversarial domain adaptation for fault diagnosis in unlabeled chemical processes

Early fault diagnosis (FD) in chemical processes can significantly enhance operational reliability and reduce energy consumption. Recently, data-driven methods based on deep learning (DL) have emerged as preferred approaches for FD. However, in complex chemical processes, models often struggle to extract invariant features from time-series data. Additionally, constructing efficient FD models with limited labeled data remains a challenge. To address these difficulties, this paper proposes a Slow Feature and Wasserstein Distance Adversarial Domain Adaptation (SWADA) method. First, a branch selection kernel fusion module based on slow feature extraction is designed to adaptively extract local deep features. These features are further learned for signal time dependencies using Long Short-Term Memory (LSTM). Second, a domain discriminator is incorporated into adversarial training, minimizing the domain shift by employing a Wasserstein distance-based metric. This promotes the extraction of domain-invariant features for classification by the feature extractor. Finally, gradient penalty is introduced to stabilize the training process during adversarial learning. Experiments on industrial three-phase flow processes (TPFP) and coke furnace processes demonstrate that the proposed method achieves superior transferable fault diagnosis performance under various operating conditions.